config: quantization_config.bits must be an integer for the Hub parser; exact value moves to bits_per_weight

config.json

@@ -124,7 +124,7 @@
   "quantization_config": {
     "quant_method": "exl3",
     "version": "1.5.1",
-    "bits": 3.21,
+    "bits": 3,
     "head_bits": 6,
     "calibration": {
       "rows": 250,
@@ -135,7 +135,8 @@
     "vision_bits": 6,
     "mtp_bits": 4,
     "head_quant": "exl3-6bit",
-    "embed_quant": "int8"
+    "embed_quant": "int8",
+    "bits_per_weight": 3.21
   },
   "attention_bias": false,
   "attention_dropout": 0.0,